Descriptions:
Jaya Gupta and Ashu Garg, partners at Foundation Capital, join the Latent Space podcast to unpack their “Context Graphs” framework — a concept they developed over late 2025 after working closely with AI-native portfolio companies struggling to get production agents to reliably execute enterprise workflows. The core insight is that current agent failures in enterprise settings are not primarily capability gaps: models lack access to the contextual reasoning embedded in human decisions — why exceptions get granted, which precedents apply, how conflicts are resolved. A context graph is the structured accumulation of those decision traces over time, becoming a company’s defensible data moat as underlying model capabilities commoditize.
The conversation draws a careful distinction between “systems of agents” — multi-agent, multi-player deployments that automate full business processes with humans in the loop — and simpler single-player chatbot interfaces. Garg, who previously ran ML infrastructure at Microsoft and seeded Databricks, brings a data infrastructure lens to how context should be stored and queried, while Gupta connects the framework to portfolio companies like Player Zero, Olive, and Factory.
The discussion also tackles harder questions: what happens when captured decision traces contradict each other (the “Swiss cheese” problem), the difference between read-path and write-path demands for a context store, and whether this framework maps onto existing concepts like AWS IAM authorization chains. Listeners building production agentic systems will find the framing of context as a first-class infrastructure layer — not a prompt engineering afterthought — practically useful.
📺 Source: Latent Space · Published February 04, 2026
🏷️ Format: Deep Dive
